This isn't necessarily true. I'm working on a local harness that doesn't do this and instead coerces everything to YAML (including tool calls) for better bucketing. Some models are indeed trained on the `<|tool_call>...<tool_call|>` token schema (or something similar—e.g. jinja), but it's vendor-specific and often times inconsistent (so you're constantly fixing calls or going back to the LLM).